Exhibit Hall | Forum 2
Purpose: The selection of non-coplanar beam orientations (e.g. gantry and couch angles) for lung cancer patients treated with IMRT is either done by protocol or manually, as a results of that, the final plan may be sub-optimal or tedious and time consuming to generate. We propose to use a deep neural network to efficiently find a superior set of beam angles that can be used for lung cancer IMRT training.
Methods: A deep supervised neural network (DNN) is trained from an algorithm known as column generation (CG), to learn and estimate the best gantry and couch angles pairs. Once trained the (DNN) is capable of generating the optimal set of beam angles in under a second, as opposed to about 20 minutes for the column generation algorithm. Using this evaluation, a set of up to 12 gantry-couch angles are selected, while limiting the number of couch rotation to five.
Results: The training loss function, MSE of the final network is at 95%. We compare the results of the predicted plans with the clinical plan available for the pair of 10 beams. To be fair, we limit the number of pairs selected for the comparison to be limited to 10 as well. The results shows that the proposed plan is 85% better than the clinical plans.
Conclusion: The proposed method can efficiently improve the quality of the treatment plan efficiently, while it is highly effective in clinical application, it is very fast, because unlike column generation calculation, the proposed method only needs to calculate the dose influence matrices for the final set of angles, which at the end can find the final treatment plan for a patient in less 5 minutes.
Funding Support, Disclosures, and Conflict of Interest: This project is supported by the National Institutes of Health (NIH)
Treatment Planning, Radiation Therapy, Lung
TH- External Beam- Photons: Treatment planning using machine learning/Knowledge Based Planning/automation